Graduate Research and Discovery Symposium (GRADS)
Predicting Optimal Error-Bounded-Lossy-Compression (EBLC) Configuration
Traditional lossless compressors can compress information to a smaller size, but don’t perform well on all data types. However, there are many domains do not require the precision of lossless compressors. It can be challenging and time consuming to determine what parameters of the compressor produce the best compression ratio or bandwidth. We propose to construct a model that predicts a high-quality configuration for the SZ error-bounded lossy compressor (EBLC). This model will aid designers in choosing an appropriate configuration given their data distribution. This poster presents and validates a modeling approach and model of the compression ratio, compression bandwidth, and decompression bandwidth for the SZ – one of the leading ELBCs. We demonstrate that the configuration of the compressor has significant impacts on these performance parameters using synthetic datasets. We further show that these impacts are predictable considering only properties of the sampled distribution. Finally, we are in the process of validating our model of optimal by comparing our models suggested parameters for real datasets from high performance computing (HPC) and intelligent transportation systems (ITS) to an exhaustive parameter sweep. We present preliminary work in this area. The model produced by this poster has broad application to these domains. It will benefit systems designers who work with quantified uncertainty who need higher performance out of existing system.
Underwood, Robert; Calhoun, Jon; and Apon, Amy, "Predicting Optimal Error-Bounded-Lossy-Compression (EBLC) Configuration" (2019). Graduate Research and Discovery Symposium (GRADS). 253.